<![CDATA[Time-series clustering remains challenging because many SOM-based studies still improve either temporal similarity measurement or winner adaptation separately, leaving limited evidence on the effect of integrating both components in a single framework. This study proposes a hybrid Self-Organizing Map (SOM) that combines Dynamic Time Warping (DTW) and a multi-Best Matching Unit (multi-BMU) mechanism for temporal pattern clustering. The proposed approach replaces Euclidean distance with DTW in representative-neuron selection and extends learning from a single BMU to a top-k winner scheme with k=3. The model was evaluated on 2,505 room-lighting time-series samples of length 25 collected over 60 weeks in an office building in Indonesia. Four models were compared across 10 independent runs: M1 (standard SOM), M2 (SOM + DTW), M3 (SOM + multi-BMU), and M4 (hybrid SOM + DTW + multi-BMU). The results show that M4 achieved the best clustering quality, with the highest mean Silhouette score (0.768300), the lowest Davies–Bouldin Index (0.4011), and the highest Calinski–Harabasz Index (233.1750). In contrast, M1 retained the lowest mean Topographic Error (0.150858). Statistical tests confirmed significant differences among models. Although the proposed model increases computational complexity compared with standard SOM, the use of a DTW window constraint helps reduce the alignment cost while preserving temporal sensitivity. The integration of DTW and multi-BMU substantially improves temporal cluster quality, although it is accompanied by reduced topological preservation and increased computational complexity.]]>
Copyrights © 2026
